See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322233711

Determining Cervical Cancer Possibility by Using Machine Learning Methods

Article · December 2017

CITATIONS READS

10 4,062

3 authors:

Muhammed Fahri Ünlerşen Kadir Sabanci

Necmettin Erbakan Üniversitesi Karamanoglu Mehmetbey Üniversitesi
36 PUBLICATIONS   31 CITATIONS    71 PUBLICATIONS   276 CITATIONS   

SEE PROFILE SEE PROFILE

Muciz Özcan
Necmettin Erbakan Üniversitesi
26 PUBLICATIONS   30 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

TÜBİTAK View project

RAC-LAB Autonomous Systems View project

All content following this page was uploaded by Muhammed Fahri Ünlerşen on 12 January 2018.

The user has requested enhancement of the downloaded file.

International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71

Determining Cervical Cancer Possibility by Using Machine

Learning Methods
Muhammed Fahri Unlersen1, Kadir Sabanci2, Muciz Özcan1
1
Department of Electrical and Electronics Engineering, Necmettin Erbakan University, Turkey
2
Departmentof Electrical and Electronics Engineering,KaramanogluMehmetbey University, Turkey

Abstract: World Health Organization (WHO) calls cancer as a generic term for a large group of diseases that
can affect any part of the body. Sometimes it could be the cause of loss of patients. Cancer mortality can be
reduced if cases are detected and treated early. In this fact, it is important to determine someone has highly
cancer risk by using a survey. In this study, a classification of patients due to their answers to a survey has been
done to determine someone who has highly cervical cancer risk. The dataset has 858 records with 33 attributes
and a biopsy result. In the dataset, number of the patients with cervical cancer diagnosis is 55 and the number of
healthy patient is 803. This dataset has been divided into two groups randomly as train and test. The train group
is 66% of the main dataset so there are 565 records in the train dataset. The rest of the dataset has been assigned
as test dataset. The test dataset has 292 records. The classification has been done by using various methods like
Multilayer Perceptron, BayesNet and k-Nearest Neighbor. Correctly classified instances and percentage, true
positive and false positive classified instances rate for each class and confusion matrix have been presented for
each method. And all of the results are discussed.
Keywords: BayesNet, Cervical Cancer, Classification, Early Diagnosis, k-NN, Multilayer Perceptron.

I. INTRODUCTION
Cancer is a leading cause of death worldwide, accounting for 8.8 million deaths in 2015. Cancer arises
from the transformation of normal cells into tumor cells in a multistage process that generally progresses from a
pre-cancerous lesion to a malignant tumor. When identified early, cancer is more likely to respond to effective
treatment and can result in a greater probability of surviving, less morbidity, and less expensive treatment.
Significant improvements can be made in the lives of cancer patients by detecting cancer early and avoiding
delays in care [1]. As in many other diseases, the existence of several screening and diagnosis methods creates a
complex ecosystem from a Computer Aided Diagnosis (CAD) system point of view. Especially in developing
countries, there are so restricted resources. Additionally, sometimes patients do not take care to routine
screening. Therefore, the most important problems during diagnosis are determination of the finest screening
plan and estimation of individual risk each patient. In most of these screening methods results have been highly
correlated with the experience of the physician and its subjective decision [2].
To reduce of unnecessary screenings, a survey could be apply to patients to determine most risky
group. So the patients could be take into a screening with a plan prepared according to the order of their cancer
risk.
In literature it is possible to find many studies about evaluation of some medical screenings. Sen et al.
(2013) in their paper have presented artificial neural network in pancreatic disease diagnosis based on a set of
symptoms. An approach to detect the various stages of pancreatic cancer affected patients has been presented in
this article. According to the manual detection procedure, using neural network detection has been more
efficient [3]. Fernandes et al. presented a regularization-based TL approach to transfer the contribution type for
each feature on linear models. In order to show its adequacy to different contexts, the proposed model-
relatedness regularize was instantiated to several learning tasks related to cervical cancer screening. It was
pointed out that positive results have been obtained [4]. In another study by Kalyankar et al., a survey of
different type of cancer such as skin cancer, breast cancer, pancreatic cancer detection using different advanced
technology such as Improved Genetic Algorithms, Artificial Neural Network, Hierarchical Clustering, Neuro-
Fuzzy system, Raman Spectra has been presented. In this study, classification performance has varied between
80.5% and 95.8% depending on the cancer type and classification [5]. In 2016, Kanimozhi et al. have presented
a study of different data mining techniques that can be employed in automated heart disease prediction systems.
Various techniques and data mining classifiers are defined in this work for efficient and effective heart disease
prediction. Each method has been provided from various studies in the literature which has used different
databases. So, the success rates have varied between 45% to 99.1% depending on classification methods and
attributes in the used databases [6]. The article that belongs Fatima et al. in 2017, has presented the comparative
analysis of different machine learning algorithms for diagnosis of different diseases such as heart disease,

www.ijlret.com 65 | Page
International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71
diabetes disease, liver disease, dengue disease and hepatitis disease. The classification performances of methods
mentioned in this study has been given in a table. The most successful method is Naive Bayes with 97% of
success rate and the worst performance is belongs to Neural Network with %70 success rate [7].
In our study a dataset created by data collected from a survey has been classified. Same classification
could be done by an expert but to improve objectivity of results, the evaluation of the screening outputs need to
be done by machine learning methods. So the classification has been done by machine learning methods like
Multilayer Perceptron (MLP), BayesNet and k-Nearest Neighbor (kNN).

II. MATERIAL AND METHODS

The dataset has been obtained from the dataset archive belongs to the University of California, Irvine.
The dataset has been collected at Hospital Universitario de Caracas in Caracas, Venezuela. The dataset
comprises demographic information, habits and historical medical records of 858 patients. The attributes in the
dataset have been presented in the Table 1 [4].

Table 1. Attribute Information

Feature Type Feature Type

Age int STDs:pelvic inflammatory disease bool

# of partners int STDs:genital herpes bool
Age of 1st intercourse int STDs:molluscumcontagiosum bool
# of pregnancies int STDs:AIDS bool

Smokes bool STDs:HIV bool

Smokes years int STDs:Hepatitis B bool
Smokes packs/year int STDs:HPV bool
Hormonal Contraceptives bool STDs: Number of diagnosis int

Hormonal Contraceptives years int STDs: Time since first diagnosis int
IUD bool STDs: Time since last diagnosis int
IUD years int Dx:Cancer bool
STDs bool Dx:CIN bool

STDs number int Dx:HPV bool

STDs:condylomatosis bool Dx bool
STDs:cervicalcondylomatosis bool Hinselmann: target variable bool
STDs:vaginalcondylomatosis bool Schiller: target variable bool

STDs:vulvo-perinealcondylomatosis bool Cytology: target variable bool

STDs:syphilis bool Biopsy: class ortarget variable bool

There are 35 attributes in the dataset. The dataset has been divided into two groups as train and test
randomly. The train dataset is 66% of the main dataset. So there are 566 records in the train dataset. The rest of
the main dataset has been assigned as the test dataset. So the test dataset has 292 records. The train dataset has
been used during training of proposed model. The obtained success rates was calculated by executing the trained
model on the test dataset.
In machine learning and statistics, classification is the problem of identifying to which of a set of
categories (sub-populations) a new observation belongs, on the basis of a training set of data containing
observations (or instances) whose category membership is known. An example would be assigning a given

www.ijlret.com 66 | Page
International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71
email into "spam" or "non-spam" classes or assigning a diagnosis to a given patient as described by observed
characteristics of the patient (gender, blood pressure, presence or absence of certain symptoms, etc.).
Classification has been being used many applications. In some of these applications, it is employed as a data
mining procedure, while in others it is employed as more detailed statistical modeling [8].
Some of classification categories have been listed as follow.
 Computer vision where medical imaging, optical character recognition and video tracking could be
grouped under,
 Speech recognition,
 Handwriting recognition,
 Biometric identification,
 Document classification,
 Internet search engines,
 Pattern recognition etc .

There are various performance indicators. In this study because of having only two class, Percentage of
Correctly Classified Instances (PCCI) has been used as performance indicators. In the following expressions,
'Positive' means that biopsy test highly recommended and vice versa. The results could be divide into four
group. They are;
 Correctly Classified Class 0 Instances also called as True Negative Class 0 (TNC0)
 Falsely Classified Class 0 Instances also called as False Positive Class 0 (FPC0)
 Falsely Classified Class 1 Instances also called as False Negative Class 1 (FNC1)
 Correctly Classified Class 1 Instances also called as True Positive Class 1 (TPC1)

These groups could be presented as in Table 2.

Table 2. Result Groups (Confusion Matrix)

Classified as
Class0 Class1
Class 0 instances that classified as Class 0 Class 0 instances that classified as Class 1
Class0
(TNC0) (FPC0)
Class 1 instances that classified as Class 0 Class 1 instances that classified as Class 1
Class1
(FNC1) (TPC1)

Number of Total True Classified Instance (TTCI) has been calculated by Equation 1.
𝑇𝑇𝐶𝐼 = 𝑇𝑁𝐶0 + 𝑇𝑃𝐶1 (1)
In this study, classification algorithms has been executed on WEKA. WEKA developed by Waikato
University in New Zealand, is an open-source data mining software with a functional graphical interface which
incorporates machine learning algorithms [9]. WEKA includes various data pre-processing, classification,
regression, clustering, association rules, and visualization tools. The algorithms can be applied on the data
cluster either directly or by calling via Java code [10, 11]. They are also suitable for developing new machine
learning algorithms.In this study tree datamining algorithm have been proposed.
k-Nearest Neighbor Algorithm: The k-NN is a supervised learning algorithm that solves classification
problems. The important point is the determination of the features of each category in advance [12]. According
to the k-NN algorithm used in the classification, based on the attributes drawn from the classification stage, the
distance of the new individual that is wanted to be classified to all previous individuals is considered and the
nearest k class is used. As a result of this process, test data belongs to the k-nearest neighbor category that has
more members in a certain class. The most important optimization problems in the k-NN method are the
identification of the number of neighbors and the method of distance calculation algorithm. In the study, the
identification of the optimum k number is performed with experiments, and the Euclidean Distance Calculations
method is used as a distance calculation method.
Euclidean calculation method has been presented in Equation 2 [13].
2 1 2
𝑑 𝑝𝑖 , 𝑝𝑗 = 𝑛𝑠=1 𝑝𝑖𝑠 + 𝑝𝑗𝑠 (2)
wherepi and pj are two different points, and the distance calculation process between them is needed. n is the
number of dimensions of the space where the pi and pj has been.

www.ijlret.com 67 | Page
International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71
Multilayer Perceptron: It is a feed forward type artificial neural network model which maps input sets
onto appropriate output sets. A multilayer perceptron (MLP) is composed of multiple layers of nodes where
each layer is connected to the next. Each node is a processing element or a neuron that has a nonlinear activation
function except the input nodes. It uses a supervised learning technique named back propagation and it is used
for training the network. The alteration of the standard linear perceptron, MLP is capable of distinguishing data
which are not linearly separable [14].
Bayes Net: It is a probabilistic graphical model and a statistical model representing a group of random
variables in addition to their conditional dependencies through a directed acyclic graph. For instance, a Bayesian
network can represent the probabilistic relations between diseases and symptoms. When the symptoms are
given, the network can calculate the probabilities of the existence of various diseases [15].

III. RESULTS AND DISCUSSION

In this study several methods have been investigated and three methods that have the best performances
has been presented. The dataset divided into two groups as training and test. All of the methods mentioned in
this study have been trained with same train dataset. Similarly, all of the methods have been tested by same test
data set. But none of the records in the test dataset presents in the train dataset.
In kNN tests, number of neighbor has been changed between from 1 to 90. For each neighbor number,
number of instances that have no cervical cancer diagnosis that have been classified as Negative (TNC0),
number of instances that have cervical cancer diagnosis that have been classified as Negative (FNC1), number
of instances that have no cervical cancer diagnosis that have been classified as Positive (FPC0), number of
instances that have cervical cancer diagnosis that have been classified as Positive (TPC1) have been obtained
and listed.
In Table 3, the most valuable 20 of 90 tests have been presented. Results of whole kNN tests have been
presented in Figure 1.

Fig 1. Results of whole kNN tests

Table 3. Results of k-NN

kNN TNC0 FNC1 FPC0 TPC1 TTCI
1 271 12 3 6 277
2 272 13 2 5 277
3 272 11 2 7 279
4 272 13 2 5 277
5 271 12 3 6 277
6 269 11 5 7 276
7 269 11 5 7 276
8 267 10 7 8 275

www.ijlret.com 68 | Page
International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71
9 267 11 7 7 274
10 268 11 6 7 275
20 268 7 6 11 279
27 267 9 7 9 276
35 267 10 7 8 275
42 267 11 7 7 274
50 267 10 7 8 275
56 267 8 7 10 277
63 266 7 8 11 277
71 266 6 8 12 278
78 266 5 8 13 279
86 266 4 8 14 280

With MLP method, it has been investigated that presenting how many neuron in the hidden layer,
provides best result. For this purpose, confusion matrix has been recorded while number of neurons in the
hidden layer of MLP has been changed from 1 to 50. In Table 4, the most valuable 15 of 50 tests have been
presented. Results of whole MLP tests have been presented in Figure 2.

Fig. 2. Results of whole MLP tests

Table 4. Results of MLP

MLP TNC0 FNC1 FPC0 TPC1 TTCI
1 265 5 9 13 278
3 265 5 9 13 278
5 265 5 9 13 278
7 265 12 9 6 271
9 265 7 9 11 276
10 265 7 9 11 276
15 265 7 9 11 276

www.ijlret.com 69 | Page
International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71
20 265 7 9 11 276
25 265 7 9 11 276
28 267 5 7 13 280
30 265 7 9 11 276
35 265 5 9 13 278
40 265 7 9 11 276
45 265 6 9 12 277
50 266 6 8 12 278

Same datasets have been investigated by Bayes Net method. The results obtained by using Bayes Net
has been presented in Table 5.

Table 5. Results of BayesNet

TNC0 FNC1 FPC0 TPC1 TTCI
Bayes Net 265 5 9 13 278

The three methods that have best results in many data mining methods like SGD, SMO, Voted
Perceptron, KStar, Multi Class Classifier Updateable, Decision Table, J48, Random Forest, Bayes Net, MLP
and kNN, have been presented. The correctly classified instances are 280, 280 and 284 by kNN, MLP and Bayes
Net respectively. The correctly classified instance percentage could be expressed for each method in the same
order as 95.89%, 95.89% and 97.26% respectively.

IV. CONCLUSION
Although as seen in results section, the best true classification result has been obtained by Bayes Net,
the problem makes the number of falsely classified instances important. While one of the objective is to make
number of correctly classified instances maximum, one of the other important objective in cancer possibility
estimation is to make number of instances classified as false positive minimum. Because number of instances
classified as false positive means that number of patients that have cancer but not warned. In this study, this
number is presented as False Negative (FNC1). For each method number of instances classified as false negative
is 4, 5 and 6 with k-NN, MLP and Bayes Net respectively. That means that false classified instance rates are
1.37%, 1.71% and 2.05% respectively. Because of being very close results in number of correctly classified
instance between Bayes Net and k-NN methods, the false negative results become more importance during
determining of the best method. So in authors' opinion, for this problem best method is determined as k-NN
method with 86 neighbor.

REFERENCES
[1] Cancer Report. February 2017 [cited WHO World Health Organization 25.09.2017]; Available from:
http://www.who.int/mediacentre/factsheets/fs297/en/.
[2] Fernandes, K., J.S. Cardoso, and J. Fernandes, Transfer Learning with Partial Observability Applied to
Cervical Cancer Screening, in Pattern Recognition and Image Analysis: 8th Iberian Conference,
IbPRIA 2017, Faro, Portugal, June 20-23, 2017, Proceedings, L.A. Alexandre, J. Salvador Sánchez,
and J.M.F. Rodrigues, Editors. 2017, Springer International Publishing: Cham. p. 243-250.
[3] Sen, T. and S. Das, An Approach to Pancreatic Cancer Detection using Artificial Neural Network, in
Proc. of the Second Intl. Conf. on Advances in Computer, Electronics and Electrical Engineering
(CEEE2013). 2013. p. 56-60.
[4] Fernandes, K., J.S. Cardoso, and J. Fernandes, Transfer Learning with Partial Observability Applied to
Cervical Cancer Screening, in Iberian Conference on Pattern Recognition and Image Analysis. 2017:
Faro, Portugal. p. 61-69.
[5] Kalyankar, M.A. and N.R. Chopde, Cancer Detection: Survey. International Journal of Advanced
Research in Computer Science and Software Engineering, 2013. 3(11): p. 4.

www.ijlret.com 70 | Page
 International Journal of Latest Research in Engineering and Technology (IJLRET)
ISSN: 2454-5031
www.ijlret.com || Volume 03 - Issue 12 || December 2017 || PP. 65-71
[6] Kanimozhi, V.A. and T. Karthikeyan, A Survey on Machine Learning Algorithms in Data Mining for
Prediction of Heart Disease. International Journal of Advanced Research in Computer and
Communication Engineering, 2016. 5(4): p. 6.
[7] Fatima, M. and M. Pasha, Survey of Machine Learning Algorithms for Disease Diagnostic. Journal of
Intelligent Learning Systems and Applications, 2017. 9(1): p. 16.
[8] Alpaydin, E., Introduction to Machine Learning. 2010: MIT Press.
[9] Witten, I.H., E. Frank, and M.A. Hall, Data mining: practical machine learning tools and techniques.
2011, London: Elsevier.
[10] Patterson, D., et al., Performance Comparison of the Data Reduction System, in Proceedings of the
SPIE Symposium on Defense and Security. 2008: Orlando, FL.
[11] Hall, M., et al., The WEKA Data Mining Software: An Update. ACM SIGKDD Explorations
Newsletter, 2009. 11(1): p. 9.
[12] Wang, J., P. Neskovic, and L.N. Cooper, Improving nearest neighbor rule with a simple adaptive
distance measure. Pattern Recognition Letters, 2007. 28(2): p. 7.
[13] Zhou, Y., Y. Li, and S. Xia, An improved KNN text classification algorithm based on clustering.
Journal of computers, 2009. 4(3): p. 8.
[14] Gardner, M.W. and S.R. Dorling, Artificial neural networks (the multilayer perceptron)—a review of
applications in the atmospheric sciences. Atmospheric Environment, 1998. 32(14): p. 2627-2636.
[15] Frank, E., M.A. Hall, and I.H. Witten, The WEKA Workbench. Online Appendix for "Data Mining:
Practical Machine Learning Tools and Techniques". Fourth ed. 2016.

www.ijlret.com 71 | Page

View publication stats